Agent Effectiveness

Agent Effectiveness is a suite of five tools that close the structural gap between what an agent knows and what it should know. Current agent tooling lets you query knowledge, check constraints, and verify before acting. What it cannot do is assess what knowledge is missing relative to a specific task, preserve reasoning state across session boundaries, or contribute institutional knowledge discovered during work. These tools address all three gaps.

The coverage assessment tool (assess_coverage) answers the question every agent should ask before starting significant work: "what don't I know?" Given a task description, it extracts implied domains and action types, queries the Engine Coverage Index for artifact statistics per domain, and identifies domains with zero or stale-only coverage. The output is not a generic health report - it is shaped by the task. A code review task will check for architecture documentation, deployment pipeline artifacts, and test coverage records. An infrastructure migration task will check for network topology documents, DNS records, and rollback procedures. Gaps near constrained areas are ranked higher risk, because missing knowledge adjacent to a declared constraint is more likely to cause damage than missing knowledge in an unconstrained domain.

The scoped context briefing tool (get_relevant_context) solves the context window problem for long-running agent sessions. During a large build, an agent needs masterplans, constraints, audit results, alerts, and decision history - far more than fits in any context window. This tool delivers a task-shaped briefing within a specified token budget, ranked by risk-if-missed rather than recency or relevance score. The ranking order prioritises block-severity constraints, then unresolved Watch alerts, then require-approval constraints, then stale-context decisions, then warnings, then fresh knowledge. High-risk items get full summaries; lower-risk items get compressed summaries; items that don't fit are counted so the agent knows what it's missing.

The contextual escalation tool (escalate_with_context) handles the class of decisions where the agent knows it's uncertain but the system can't evaluate why. Unlike verify_before_acting (which evaluates actions against known constraints), escalation packages the agent's full reasoning state - evidence gathered, options considered with pros and cons, a recommended action, and the reasoning behind the uncertainty - and routes it to a human reviewer. When the human responds, the guidance is incorporated into the constraint surface for the session. If no human responds within the SLA, the agent receives a fallback signal rather than blocking indefinitely.

The decision checkpoint tool (checkpoint_decision_state) preserves agent reasoning state across session boundaries. When an agent fails mid-build or hits an unrecoverable error, its decisions, assumptions, and open questions normally evaporate. Checkpoints create receipted snapshots that a new agent instance (or the same agent in a new session) can resume from. Intermediate checkpoints are compacted when a session succeeds; failed sessions retain all checkpoints for forensic analysis.

The capability discovery tool (get_platform_capabilities) provides a machine-queryable manifest of all available tools, their required trust tiers, and credit costs. This enables agent-to-service evaluation without reading documentation - an agent can programmatically determine which tools are available, what they cost, and whether its current trust tier permits their use.